How AI companions work — a plain-language guide for parents

The three parts of every companion

Strip away the marketing and every AI companion — app, plush toy, or robot — is three systems working in a loop. Ears: a microphone plus speech recognition that turns sound into text. Brain: a language model that decides what to say back. Voice: text-to-speech that says it aloud. Everything else — the character, the personality, the safety promises — is configuration wrapped around that loop.

The single most important question a parent can ask is where each part runs. On-device processing means the audio never leaves your home. Cloud processing means your child's voice travels to a server, is processed, and (in the worst products) is stored. Reputable makers publish this split clearly; evasive answers are themselves an answer.

What the 'brain' actually does

A language model is a prediction engine: given the conversation so far, it predicts a likely next reply, word by word. It has no goals, no feelings, and no understanding in the human sense — but it is extraordinarily good at sounding like it does. That is precisely why children take to it, and precisely why guardrails matter.

Good child products do not connect a raw model to a child. They add layers: a system prompt that fixes the persona and tone, topic filters that catch unsafe requests before the model sees them, response filters that check the answer before it is spoken, and escalation rules — 'that is a great question for your mum or dad' — for anything sensitive. When you evaluate a product, you are really evaluating these layers, not the model underneath.

The data questions that matter

Ask five things. One: is audio processed on-device or in the cloud? Two: are transcripts stored, for how long, and can you delete them? Three: is your child's data used to train models — and is that off by default? Four: can the companion make purchases or open links? Five: what exactly does the parent dashboard show you, and what does it show the company?

A trustworthy answer pattern looks like: minimal collection, short retention, no advertising, no third-party sale, training opt-in rather than opt-out, and a parent view that is at least as informative as the company's own view. Regulation such as COPPA in the US and GDPR-K principles in Europe set floors, not ceilings — the best products go beyond them.

How RAFEEQI approaches this

We are designing RAFEEQI so the answers to those five questions are boring: as much on-device as hardware allows, parent-visible summaries instead of secret transcripts, no ads ever, and safety filters that run before and after the model. Our demo today is fully scripted — no live model — because we would rather show an honest prototype than ship an unguarded brain. This guide is the standard we expect to be held to.